University of Southern California

Reinforcement Learning for the Lux AI Challenge

Abstract

The Lux AI Challenge is a two-player perfect information game in which competitors control workers who collect resources and build the largest cities possible. In this project, we experiment with a variety of approaches to the challenge: traditional rule-based agents based on human intuition, deep reinforcement learning agents trained from scratch, evolution strategy agents trained from scratch and from warm starts, and imitation learning agents trained from other high-performing players. Of our approaches, we achieve the best performance with a hybrid approach consisting of an imitation learning agent fine-tuned with an evolution strategy, placing 89th out of 1,122 teams on the competition leaderboard (top 8%) as of writing.